Non-orthogonal multiple access (NOMA) is a potential candidate to further enhance the spectrum utilization efficiency in beyond fifth-generation (B5G) standards. However, there has been little attention on the quantification of the delay-limited performance of downlink NOMA systems. In this paper, we analyze the performance of a two-user downlink NOMA system over generalized {alpha}-{mu} fading in terms of delay violation probability (DVP) and effective rate (ER). In particular, we derive an analytical expression for an upper bound on the DVP and we derive the exact sum ER of the downlink NOMA system. We also derive analytical expressions for high and low signal-to-noise ratio (SNR) approximations to the sum ER, as well as a fundamental upper bound on the sum ER which represents the ergodic sum-rate for the downlink NOMA system. We also analyze the sum ER of a corresponding time-division-multiplexed orthogonal multiple access (OMA) system. Our results show that while NOMA consistently outperforms OMA over the practical SNR range, the relative gain becomes smaller in more severe fading conditions, and is also smaller in the presence a more strict delay quality-of-service (QoS) constraint.